Developing an Artificial Intelligence Model for Estimating the Compressive Strength of Concrete Based on Combined Non-Destructive Test Results

Abstract

The most important mechanical property of concrete for design and evaluation is its compressive strength, which, for new structures, is typically determined by breaking cylindrical or cubic concrete specimens under uniaxial load in controlled laboratory conditions. The accurate and quick evaluation of the compressive strength of concrete in existing structures has been of primary concern to engineers and researchers for decades. The evaluation of concrete compressive strength for existing reinforced concrete (RC) structures is currently done mainly by destructive testing (DT) of extracted core samples which can be both costly and time consuming. To be representative of the in-situ concrete, a relatively large number of samples are needed which is often impractical or unfeasible. Non-destructive test (NDT) methods have been developed to provide indirect measurements that can be correlated to concrete strength. Among them, the SonReb method, which combines ultrasonic pulse velocity (UPV) readings with rebound hammer (RN) measurements, has shown promising results. This approach can reduce the number of concrete cylinders or cores that are needed, saving time and money for building owners. However, despite its advantages, the need to develop calibration curves for each project presents a notable limitation of all NDT methods. This study is the first comprehensive investigation in which a Machine Learning (ML)-based model is developed using the SonReb method for evaluating the actual in-place compressive strength of concrete in existing RC structures using only NDT measurements, without the need for new calibration curves. In this study, ML techniques are applied to improve the SonReb method for practical use in four phases: 1) developing a preliminary ML-based model and graphical user interface (GUI) that converts UPV and RN values to estimates of concrete compressive strength; 2) evaluating the initial ML model performance against three case studies to assess the opportunities and limitations of the proposed approach in practice; 3) comparing the performance of alternative ML algorithms and adding input parameters to account for specimen geometry to improve prediction accuracy; and 4) assessing the uncertainty and prediction intervals of the new ML-based model versus conventional regression models. The ML-based SonReb approach presented in this thesis was, in general, able to provide reasonable approximations of concrete compressive strengths for various concrete mix designs, with mean absolute percent errors (MAPEs) within 10-15%, without calibration. Some discrepancies were also noted, which highlight the role of engineering judgment in proper interpretation of results and use of modern and conventional tools. The results of this study suggest that although artificial intelligence (AI) has the potential to improve predictions of the in-place compressive strength of RC structures using simple and fast NDT methods in certain cases, challenges related to data collection can affect model performance. A road map for improved data collection is proposed that can be used for future development of a reliable and robust ML prediction model for concrete compressive strength of existing structures.